Vacancies in DARC

Machine learning driven balance relationships for next generation data assimilation systems
Lead Supervisor: Ross Bannister, Univ of Reading, Dept of Meteorology and National Centre for Earth Observation Email: r.n.bannister@reading.ac.uk
Co-supervisors: Christopher Thomas, Met Office; Varun Ojha, Univ of Reading, Dept of Computer Science; Hong Wei, Univ of Reading, Dept of Computer Science

Are you fascinated by the complex models and systems that are used to produce weather forecasts? Are you a physicist/engineer/mathematician/meteorologist/computer scientist who would like to work towards a PhD with the Met Office in this important area of scientific endeavour?

The Met Office is experimenting with very high resolution models covering regional domains (e.g. a ~100m grid length model over London, UK). Models with such high resolution are right at the leading edge of weather forecasting, but are able to produce useful forecasts only if they are initialised close to the current weather and in a way that is consistent with the model dynamics. Even with modern observing systems, including radar and satellites, this process is currently far from optimal and so a lot of fundamental research is required to improve it.

Integrating dynamical models and observations is a formal process called data assimilation, which makes up a significant portion of the numerical weather prediction computation burden. Data assimilation attempts to use the latest (but limited) observational information to infer the initial conditions of the model so that useful forecasts can be made. This process projects reality (as observed) to the model’s state, but this projection must not be allowed to excite inappropriate instabilities in the ensuing forecasts, so the assimilation must also be done in a way that leaves the model balanced.

Ensuring appropriately balanced forecasts requires statistical information about how different components of the model’s field should be made to co-vary as the observations are assimilated. Such information is highly dependent on the flow itself and is not easily available in real time. The hypothesis of this PhD project is that machine learning techniques (like neural networks) can be used to diagnose useful covariances between different positions and variables in the model state so that unwanted instabilities can be avoided. The challenges will be to understand the covariances that should be imposed, to select and apply an appropriate machine learning method, and then to show how these covariances can impact the data assimilation and subsequent forecasts.

This is a stimulating task at the intersection of meteorology, data science, and computer science.

Training opportunities:
As a student on this project, you will be supervised by and work with academics at the University of Reading and the Met Office. Training will be offered in data assimilation, machine learning, meteorology, and programming in the form of masters-level modules and summer schools. The studentship is supported by SCENARIO and the Met Office and as such you will spend some of the time working at the Met Office.

The student will primarily be a member of the Department of Meteorology at the University of Reading, but also part of the Department of Computer Science, the Data Assimilation Research Centre (DARC), and the National Centre for Earth Observation (NCEO) and so will have access to facilities offered by these organisations.

Student profile:
The ideal student would have an excellent degree in physics, engineering, mathematics, meteorology, computer science, or another highly quantitative and analytical discipline. The student will be willing to learn the mechanics of data assimilation and machine learning and show creative and innovative thinking to demonstrate new knowledge in this challenging area. The project will require complex coding, including with the Met Office’s new JEDI-based system.

Funding particulars:
SCENARIO is the primary potential sponsor of this project, and, should it go ahead, the studentship will have CASE sponsorship from the Met Office.  Support from the NCEO is also being sought.  Important note: this project has been approved to be advertised, but funding is not yet guaranteed (about 1/3 chance of final funding).

How to apply:
To apply, please visit the SCENARIO web page, where instructions will be given in due course.

research.reading.ac.uk/scenario

---

Treatment of model bias in marine biogeochemistry forecasting

Lead Supervisor: Alison M. Fowler, University of Reading, Department of Meteorology and National Centre for Earth Observation (NCEO) a.m.fowler@reading.ac.uk
Co-supervisors: Jozef Skákala, Plymouth Marine Laboratory and NCEO; Amos S. Lawless, University of Reading, Department of Mathematics and NCEO

The monitoring and forecasting of the marine biogeochemistry in the shelf seas is essential for understanding the present and future health of our seas and its many associated environmental, economic and societal impacts. The forecasting includes long-range forecasts addressing the impact of climate change on ocean biological production and acidification, as well as shorter time-scale forecasts predicting sudden dangerous events, such as harmful blooms, or hypoxia. To enable numerical models of the marine biogeochemistry to stay in line with the true underlying system a technique known as data assimilation is used to systematically blend the model with observations made from a myriad of instruments. One rich source of observations comes from satellite derived sea surface chlorophyll concentrations (see inset figure), a proxy of how much life there is in the ocean.

Data assimilation algorithms are based on mathematical principles that make approximations about the errors in both the model and observations. One of the most fundamental approximations is that both are unbiased estimates of the true underlying system. Unfortunately, in many applications biases in the model remain significant and so the data assimilation is suboptimal. The magnitude of the biases in marine biogeochemistry are particularly large and a known limitation in the forecast skill. Different approaches to treating bias within the assimilation exist, but in order to apply these techniques an estimate of the bias or its statistics are needed. This is particularly challenging given that the biases are likely to have high variability in space and time. Utilising machine learning techniques, this project aims to develop parameterisations of model bias that allow for their estimation from the model and observation data available. Different state-of-the-art techniques for the correction of the bias during the assimilation will then be applied to idealised models in which their sensitivity to the accuracy of the estimated bias is assessed, before then being applied to an operational model of the marine biogeochemistry in the shelf seas surrounding North West Europe.

Training opportunities:
This studentship is a joint project with the Plymouth Marine Laboratory (PML). The student will have the opportunity to spend time working at PML over the lifetime of the project. The student will also have the opportunity to attend ECMWF training courses on data assimilation and advanced training courses at Reading organized by the Data Assimilation Research Centre and the National Centre for Earth Observation.

Student profile:
This project would be suitable for students with a good honours degree in a subject with strong mathematical content and programming experience.

https://research.reading.ac.uk/scenario/

---

New approaches to ocean state analysis for climate and forecasting applications.

Lead Supervisor: Keith Haines, University of Reading, Department of Meteorology Email: k.haines@reading.ac.uk
Co-supervisors: Daniel Lea, UK Met Office; Matthew Martin, UK Met Office

Reconstructing present-day and past ocean temperature salinity and circulation states is critical both to making long-range weather and climate forecasts and for understanding how the ocean has responded over the last century to imbalances in the Earth’s energy budget due to global warming. However most observations of the oceans are only available at the surface and widespread subsurface observations have only become available in the last 15 years. If we could understand better how to use surface and subsurface ocean observations to complement each other more effectively we could greatly improve reconstructions of ocean phenomena such as the Gulf stream and the Antarctic circumpolar current in the southern ocean, and the eddies and vortices that develop around strong currents and fronts between colder and warmer waters. We would also be able to reconstruct better records of changes in ocean heat and salinity distributions using the sparse ocean observations from the past which could help us understand how the oceans have warmed due to climate changes over the past century.

In this project you will work with Reading and Met Office staff to develop and compare a range of new methods for reconstructing the ocean state using satellite observations of sea level and sea surface temperatures, along with profiles of subsurface temperature and salinity, T(z), S(z), observed by autonomous “Argo” robotic profilers (https://argo.ucsd.edu/). Satellite altimeters provide a record of global sea level starting in 1992, with areas of higher sea level generally indicating warmer waters below, but they do not give information on the vertical heat distribution. In-water profile measurements give detail of the subsurface ocean structure but are much more limited in spatial extent. You will investigate how best to combine these data, as well as the impact of the reconstructed ocean states on computer forecasts of changes in ocean temperatures and associated weather. You will also investigate how to use results obtained from the greater volume of current surface and subsurface ocean observations in order to better reconstruct ocean states in the past when less data were available.

This project has a strong focus on data analysis and you will develop new methods of combining observations with modelled ocean data which is required for making all ocean and atmospheric forecasts. These skills are of great importance for developing improved forecasting methods and are in great demand at operational forecasting centres such as the Met Office. You will learn about the role of the oceans in both marine and weather forecasting. There is also growing interest in using novel Machine Learning methods for understanding large climate datasets and you will explore how these methods can be brought into use for analysis of the ocean circulation.

Training opportunities:
The student will benefit from in-house training in oceanography and especially data assimilation methods where the Reading Meteorology and Maths departments have a world-renowned centre. The Computer science department is also part of our university school providing training in novel data science methods.

The student will have regular meetings with Met Office staff and will spend several periods working in the Met Office marine forecasting group in Exeter. They will also benefit from extensive research training opportunities through the SCENARIO Doctoral Training program.

Student profile:
We are looking for a strongly motivated individual with an interest in the oceans and climate and in data analysis, and a background in the physical sciences, meteorology or oceanography, applied mathematics or computational sciences to take on this project. We will provide training on modelling and computer programming to motivated candidates as needed, however confidence in solving numerical problems computationally would be an advantage.

Funding particulars:
This project has CASE sponsorship from the UK Met Office.

https://research.reading.ac.uk/scenario/

---

Maximising the value of observational data in ensemble data assimilation for hazardous weather prediction

Lead Supervisor: Sarah L Dance, University of Reading/National Centre for Earth Observation Email: s.l.dance@reading.ac.uk

Co-supervisors: Joanne Waller, Met Office

In a changing climate, an improved ability to forecast hazardous weather is key to the management of risk for society. In weather forecasting systems, large numerical models solve nonlinear equations describing physical processes in the atmosphere. Data assimilation is routinely used to improve weather predictions by combining billions of variables from these numerical model simulations with millions of observations of the atmosphere. Data assimilation can be thought of as a machine learning or mathematical optimization approach, where a cost function is minimized. The cost function is essentially a weighted measure of the distance from forecast states (numerical simulations) and the available observations over a fixed time window, weighted by the uncertainties (error statistics) in the data. Thus, weather forecast accuracy relies on accurate estimates of the uncertainty in weather observations. However, less than 5% of some key observation-types are assimilated, in part because these uncertainties cannot be properly quantified and accounted for. Thus, a key research question is how to characterize and treat observation uncertainty in an assimilation system. This project will investigate mathematical methods to approximate observation uncertainty that preserve observation information content while being sufficiently efficient for practical use in operational weather prediction.

 

 

 

 

 

 

 

 

In recent years, there have been great improvements in forecasting high impact weather events such as intense rainfall leading to flooding, via the use of very fine (1.5km) grid-spacings. It is important to be able to predict precise locations and timings for hazardous weather events to enable appropriate mitigating actions (such as deployment of temporary flood barriers). However, hazardous weather predictions, such as forecasts of intense summer rainfall, are sensitive to the detailed starting conditions. We address this in two ways:

• Using dense observation data from remote sensing instruments (e.g., satellites and ground based-radar) providing detailed information about the current state of the atmosphere. We need good estimates of the spatial relationships for the uncertainties in these observations to make best use of them for weather forecasting.
• Using novel ensemble methods. These methods use multiple forecasts driven from slightly different starting conditions, to explicitly represent the flow-dependent nature of the forecast uncertainty.

The interactions between the flow dependent forecast uncertainty from the ensemble, and new approaches to estimate observation uncertainties are not well understood. In the project, theoretical investigation and simplified model experiments will be used to provide valuable insights with potential to trial the best techniques with real data in the Met Office operational system. The project is flexibly designed to allow the student to tailor the project towards to their own strengths and interests, whether that is the development of mathematical theory, numerical implementation or physical insights into the sources of observation uncertainty. This will be facilitated by the use of a new software package (JEDI) that gives a choice of model complexity, from simple idealized models that run quickly on a laptop to the full operational numerical weather prediction system requiring a large database of observations and parallel computations.

The student will join a supportive research team in Reading’s Data Assimilation Research Centre and be affiliated to the National Centre for Earth Observation (NCEO), giving wider opportunities for training and networking. The project is a collaboration of the University of Reading with the Met Office, the UK’s national weather service. The researcher will be supported by advice from Met Office experts throughout the project, and access to Met Office software and observation data while on a placement. The research is expected to increase the proportion of high-resolution satellite data assimilated, leading to better forecasts of high impact weather and better value for money for investments in in satellite data.

Training opportunities:
The student will receive full academic, technical and transferable skills training through Masters level courses, short training courses and summer schools. The student will be given the opportunity to undertake a three-month placement at the Met Office (in Reading, Exeter or via online participation).

Student profile:
This project would be suitable for highly motivated students with a degree in mathematics, physics or another scientific or engineering discipline with a high mathematical content. Previous computer programming experience is desirable but not essential, as training can be provided.

Funding particulars:
This project is co-funded by the NERC SCENARIO DTP and the National Centre for Earth Observation (TBC!). It has additional co-sponsorship from the UK Met Office in the form of a CASE award. This will supply an additional £1000 per annum to the Research Training and Support Grant for three years and also funds travel and subsistence for the student to undertake a 3-month placement at the Met Office.

References:

 

• Data assimilation: The secret to better weather forecasts https://youtu.be/YPAWYjPf_Pk (non-technical video)
• Tabeart, J. M. (2019) On the treatment of correlated observation errors in data assimilation. PhD thesis, University of Reading https://doi.org/10.48683/1926.00088830

https://research.reading.ac.uk/scenario/

---